Video Friday: Meet Google DeepMind’s Gemini Robotics 2

We have spent decades trying to teach machines how to move by writing complex, rigid lines of code for every possible joint rotation. It has been slow, expensive, and frankly, quite fragile. But the latest updates coming out of Google DeepMind with Gemini Robotics 2 suggest the industry is finally hitting a pivot point. We are moving from the era of 'scripted motion' to 'generalized intelligence' in the physical world.

The End of the Script

For a long time, the barrier to useful robotics wasn't just the hardware costs; it was the lack of adaptability. If you programmed a robot to pick up a red ball and the ball was slightly more orange that day, or if the lighting shifted, the system would often fail. DeepMind is trying to solve this by layering their Gemini model directly into the robotic control stack.

What we are seeing now is whole-body control. Instead of treating a robot arm and a robot leg as separate systems that need to be synchronized through manual math, Gemini 2 treats the entire machine as a singular, intelligent entity. This allows for what they are calling 'adaptable' movement. To a builder, this means you might not have to hard-code for every edge case anymore. The model handles the physics of the real world because it has ingested enough data to understand intent, not just coordinates.

Dexterity and the 'Shadow' Problem

One of the more interesting developments in recent robotics research, specifically out of places like the General Robotics Lab, involves robots learning through abstractions—like shadows and reflections. While it sounds like a philosophical experiment, it is actually a clever way to solve for spatial awareness. Humans use their shadows to judge distance and posture without even thinking about it. Giving a 21-degree-of-freedom hand that same level of self-model awareness is a massive leap for dexterity.

For those of us building in the AI space, this is a reminder that the data we feed these models doesn't always have to be direct sensor input. Sometimes, the indirect data—the 'shadow' of the action—is what provides the most robust training ground for fine motor skills.

Form Factor: Humanoids vs. Functionality

We are seeing a massive divergence in how people think robots should look. On one hand, you have the sleek, bipedal humanoid push from companies like Agility and Unitree. They want machines that look like us so they can navigate worlds built for us. On the other hand, you have companies like Hello Robot with their Stretch 4.0—a one-armed, three-wheeled pole that looks nothing like a human but excels at navigating a cluttered living room.

As a founder, you have to ask yourself: are you solving for the aesthetic of the future, or the utility of the present? The humanoid form factor is impressive, but it’s a nightmare to repair and balance. A three-wheeled robot is boring, but it’s stable and significantly cheaper to deploy. The 'cool factor' of a bipedal robot often masks the reality that, for most warehouse or home tasks, four wheels and a simple mast are more efficient than two legs and a spine.

The Repairability Crisis

There is a quiet concern brewing in the hardware community that I think needs more sunlight: repairability. When you look at the complex hydraulics and carbon-fiber skeletons of the latest Unitree or Agility bots, you have to wonder what happens when a sensor fails or a joint snaps. If we are moving toward a world of autonomous agents, they cannot be 'disposable' high-tech toys.

For builders, the opportunity here isn't just in the AI or the motors; it's in the maintenance infrastructure. The first company that builds a high-performance robot that can be fixed with a basic toolkit and off-the-shelf parts is going to have a massive advantage over the closed-loop, proprietary giants.

Why This Matters for Builders

If you are in the crypto or AI space, you might feel like robotics is a different world. It isn't. The same large language models (LLMs) we use to generate code or write copy are now being used to translate human intent into physical torque. We are seeing a convergence where the 'brain' of the AI is finally getting a body that isn't just a chatbot window.

The takeaway for founders is clear: hardware is becoming a commodity, and the 'intelligence layer' is the new moat. If you can build a system that allows a robot to navigate a Mars-like environment—as the RAD Lab at USC is doing with NASA—using generalized learning rather than specific terrain mapping, you’ve built something that can scale to any environment on Earth.

The Long View

We are still a way off from robots doing our laundry in real-time without a 5x speed-up on the video. The 'Video Friday' highlights show progress, but they also show the gaps. We see robots folding shirts, but the footage is often manipulated to look faster than it is. We see robots playing badminton, but on smaller courts designed to make them look more capable.

Don't be fooled by the hype, but don't ignore the momentum. The transition from 'if-then' programming to 'neural-net' physical control is the most significant change in robotics since the invention of the servo motor. The builders who win this decade won't be the ones making the shiniest robots; they’ll be the ones making the most adaptable brains for the machines we already have.

The goal isn't to make a robot that mimics a human; it's to make a robot that understands the world well enough to be useful without human intervention.

We are getting closer to that reality, one Gemini update at a time. Just make sure you bring a wrench, because these things are still going to break.

Read the original at IEEE Robotics →